Millions of abdominal and chest CT scans are taken every year for reasons unrelated to bone, yet they image the lower spine in 3D. Software can outline each vertebra automatically and read its bone density from these routine scans. On a fracture-rich cohort, that density identifies patients with vertebral fractures far better than the standard dual-energy X-ray test (Löffler et al.). What these tools report, though, is a single material property compared against a fixed cutoff. Two patients with the same density can face very different risk: one is small and light with narrow vertebrae, the other is heavier and has wide ones. Environmental toxicology handles the analogous problem with the hazard quotient. It divides an estimated exposure, scaled by body weight, by a tolerance threshold, so a value above 1 means "exposure exceeds what the system can tolerate" regardless of units (Ye et al.).
We propose computing a per-vertebra quotient from a single routine CT. The numerator is the estimated compressive load on the vertebra, taken from body mass above that spinal level, with body mass itself estimated from the scan when the record lacks it. The denominator is the vertebra's estimated failure strength, derived from automatic segmentation. The main question is whether this quotient, with a threshold fixed in advance and never fitted to any hospital's data, keeps its sensitivity and specificity better across hospitals, scanners and contrast protocols than a density cutoff or a regression model fitted at one site. We answer it with random-effects meta-analysis: pooling per-hospital results while modeling genuine between-hospital variation, not just sampling noise. The decisive quantity is the prediction interval, the range of performance a new, unseen hospital should expect.
Core Claim and What Is New
Load-to-strength ratios have a history in bone biomechanics outside the papers indexed here, so we do not claim the ratio itself. The file contains three separate threads. Opportunistic spine work reports density (Löffler et al.; Ramschütz et al.) or simulated failure load (Rayudu et al.), but never normalizes by the patient's own loading. Hazard quotients appear only for dietary and environmental toxins. Random-effects pooling appears for AI diagnostics in other organs (AI prenatal ultrasound meta-analysis) and for spine surgery risk factors (Luo et al.), but not for opportunistic spine CT.
What is new is the combination of three things. Both halves of the quotient come from one routine scan. The threshold is anchored physically rather than fitted. Its transportability across sites is measured by between-site heterogeneity. The claim we would defend is narrow. At a fixed, pre-registered threshold, the quotient's site-level sensitivity varies less across hospitals than that of the fixed 160 mg/cm³ integral-density cutoff from Löffler et al.. Its patient-level discrimination is no worse than density alone. If a density-plus-body-size regression transported equally well, the quotient would reduce to a relabeling, and we would report that.
Research Program
The denominator. A segmentation network localizes and labels T11–L4, then separates the trabecular core from the cortical shell. We train it in the style of the VerSe benchmark (Liebl et al.). Density is converted from scanner intensity using asynchronous calibration, as validated across 14 cadaver protocols in Ramschütz et al.. An anatomy-guided classifier detects the contrast phase and applies linear correction, following Rühling et al. (2021) and Toelly et al..
We compute strength two ways. The cheap version is trabecular density times the mid-body cross-sectional area measured from the mask. The expensive version is finite element analysis, a physics simulation that meshes the segmented vertebra, maps density to stiffness, and finds the compressive force at which it fails, following Rayudu et al.. Their authors named manual segmentation as the bottleneck that our pipeline removes.
The numerator. For level $\ell$ of patient $j$, the quotient is
$$Q_{j\ell} = \frac{k \, f_\ell \, m_j g}{S_{j\ell}},$$
The symbols are:
- $m_j$ is body mass.
- $g$ is gravitational acceleration.
- $f_\ell$ is the standard anthropometric fraction of body mass carried above level $\ell$.
- $k$ is a fixed multiplier representing a reference daily activity, such as forward bending.
- $S_{j\ell}$ is the estimated failure strength.
Body mass is often missing from imaging archives and from public datasets. We therefore estimate it from the scan itself, using body cross-sectional areas and soft-tissue composition, and validate the estimate against recorded weight where records exist.
The threshold and its uncertainty. The multiplier $k$ does not change patient rankings. It only sets where "1" falls. This is exactly the hazard-quotient weakness that Ye et al. flag, where the reference dose for cerium came from pregnant mice and its relevance to children is uncertain. We handle it the same way that literature should. We pre-register $k$ from biomechanical convention before seeing outcome data, and we report how conclusions change over a plausible range of $k$. We then propagate uncertainty from calibration error and body-mass estimation error, turning the point quotient into a per-vertebra probability that $Q > 1$.
The meta-analytic evaluation. Each participating hospital, or each scanner model within a hospital, is one unit. Per unit we compute two things: the difference in patient-level AUC between the quotient and density alone, and the sensitivity and specificity at the fixed thresholds. We pool them with random effects, estimated by restricted maximum likelihood with Hartung–Knapp intervals. Weights are $1/(\sigma_i^2 + \tau^2)$, where $\sigma_i^2$ is unit $i$'s sampling variance and $\tau^2$ is the between-unit variance. A meta-regression then asks whether contrast phase, tube voltage, vendor, or slice thickness explain $\tau^2$.
Baselines
The naive reference is a model using age and sex only, fitted at the development site and applied unchanged elsewhere. It shows how much any imaging measure adds.
The strongest existing method is automated trabecular or integral density with the fixed 160 mg/cm³ cutoff, the pipeline of Löffler et al.. That paper reports an AUC of 0.885 for trabecular density identifying prevalent vertebral fractures in 192 patients, and 84% sensitivity with 78% specificity at the integral cutoff. That cohort had a 77% fracture prevalence, so we do not expect to reproduce these numbers on lower-prevalence data. We rerun the baseline ourselves on the identical scans, segmentations and calibration.
The second strong competitor is failure load from finite element analysis alone, without load normalization. Rayudu et al. report 4,705 ± 1,238 N for healthy and 4,010 ± 1,297 N for fractured vertebrae (p = 0.026) in 18 patients. No reported AUC was found on Lacuna for failure load as a fracture classifier.
The most dangerous competitor is a logistic regression on density, cross-sectional area and body mass as separate predictors. It is fitted at the development site with the same data and the same hyperparameter budget as any tuning we do on the quotient, then frozen for transport.
What counts as beating the baselines. To beat density we need two things. The pooled AUC difference must be non-negative, with its lower 95% bound above −0.02. The prediction interval for site-level sensitivity at the fixed threshold must be narrower than density's, with $\tau^2$ smaller in a paired comparison. To beat the frozen regression we need the same transport advantage with no loss in pooled AUC. Matching it on both counts would leave the quotient with interpretability only, and we would say so.
Ablations. The first holds the numerator at a population-average load, which reduces the quotient to strength alone and isolates the contribution of patient-specific load. The second swaps CT-estimated mass for recorded weight, which isolates body-mass estimation error. The third swaps finite element strength for density times area, which tells us whether the simulation earns its compute. The fourth pools with a fixed-effect model to show how much confidence would be overstated.
Why not just…?
Why not just use density with a fixed cutoff, since it already reaches AUC 0.885? Because that number comes from one high-prevalence university cohort, and density readings shift with protocol. The uncorrected portal-venous error reported by Rühling et al. (2021) is 18.7 mg/ml, and cadaver biases span −12.66 to 8.00 mg/cm³ (Ramschütz et al.). Shifts of that size move patients across a fixed cutoff. The quotient does not cancel that bias, since its denominator inherits it. The paired comparison of $\tau^2$ between the two measures will show whether normalizing by load and size makes misclassification more or less stable. If density's between-site variation is equal or smaller, density wins.
Why not just fit a regression with density, vertebral size and weight? This is a real threat. With enough data, a flexible model will match or beat any fixed ratio on in-site discrimination. Our wager is that a physically structured ratio with no fitted threshold transports better than coefficients tuned to one hospital. The frozen-regression baseline tests exactly this. If the regression transports as well, the paper's conclusion becomes "use the regression and report its prediction interval."
Why not just run finite element failure load and skip the numerator? Failure load already folds in size and density. The population-average-load ablation isolates what patient-specific load adds. If that ablation matches the full quotient, body mass carries no information beyond strength, and the numerator should be dropped.
Why not just pool with a fixed-effect model? Vendors, reconstruction kernels and contrast phases differ across sites (Liebl et al.). Assuming a single true effect would produce falsely narrow intervals, the overconfidence random-effects pooling is used to avoid in Luo et al.. More importantly, a fixed-effect model has no prediction interval, and that interval is our primary outcome. The cost of random effects is that $\tau^2$ is imprecise when there are few sites. We target at least eight units and report the uncertainty in $\tau^2$ openly.
Proposed Experiments
Segmentation. We first confirm that the segmenter is competitive on VerSe 2020, where the challenge's top entry reached a 96.6% identification rate and 91.72% Dice (Liebl et al.). We then measure how often a label shifted by one level changes the quotient's classification. The numerator depends on level, so an off-by-one label (T12 read as L1, for example) matters here in a way it does not for density.
Development cohort. We use the fracture-graded dataset of 160 scans with 1,725 vertebrae, Genant grades and lumbar density (vertebral fracture grading dataset), plus VerSe's multi-vendor scans. This supports development and within-dataset evaluation. No segmentation Dice is reported for the fracture-graded dataset in the file.
Leakage from fractured vertebrae. A fractured vertebra has altered geometry and density, which would flatter any image measure. We therefore define the patient-level score from unfractured levels only.
Transport test. Neither public dataset records body weight, and a genuine transport test needs many sites. The core experiment therefore uses retrospective routine CT from partner hospitals, with recorded weight, contrast phase and scanner parameters. The primary endpoint is prevalent fracture. A secondary endpoint is incident fracture on follow-up CTs already in the archive, which addresses the cross-sectional limitation of Löffler et al..
Decisive checks. Five results settle the proposal:
- The paired between-site variance of sensitivity at the fixed threshold, quotient versus density.
- The pooled AUC difference with its prediction interval.
- The frozen-regression comparison.
- The population-average-load ablation.
- The meta-regression showing whether contrast phase still explains heterogeneity after phase correction.
If checks 1 and 3 both fail, the quotient offers nothing beyond relabeling, and we will report that result.
References
- Automatic opportunistic osteoporosis screening in routine CT: improved prediction of patients with prevalent vertebral fractures compared to DXA
- Occurrence and Exposure Assessment of Rare Earth Elements in Zhejiang Province, China
- Automated extraction of vertebral bone mineral density from imaging with various scan parameters: a cadaver study with correlation to quantitative computed tomography
- Predicting Vertebral Bone Strength Using Finite Element Analysis for Opportunistic Osteoporosis Screening in Routine Multidetector Computed Tomography Scans—A Feasibility Study
- Artificial intelligence-enabled prenatal ultrasound for the detection of fetal cardiac abnormalities: a systematic review and meta-analysis
- Risk factors of epidural hematoma in patients undergoing spinal surgery: a meta-analysis of 29 cohort studies
- VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images
- Automated detection of the contrast phase in MDCT by an artificial neural network improves the accuracy of opportunistic bone mineral density measurements
- Influence of Contrast Media on Bone Mineral Density (BMD) Measurements from Routine Contrast-Enhanced MDCT Datasets using a Phantom-less BMD Measurement Tool
- A Vertebral Segmentation Dataset with Fracture Grading